Complete Fuzzy Knowledge Representation With Knowledge Graph Embedding.

Area of Science:

• Artificial Intelligence
• Data Science
• Machine Learning

Background:

• Fuzzy knowledge graphs (KGs) represent uncertain facts, but existing methods struggle with noise amplification and interpretability.
• Current approaches lack generality and often entangle semantic and reliability signals.

Purpose of the Study:

• To propose FGF-GAT, a general encoder-decoder framework for fine-grained fuzzy knowledge graph embedding (KGE).
• To address limitations in neighborhood aggregation, signal entanglement, and restricted generality in existing fuzzy KGE methods.

Main Methods:

• FGF-GAT utilizes a learnable neuro-fuzzy reasoning module for fine-grained fuzziness modeling, transforming memberships into reliability-aware signals.
• An ANFIS-inspired fuzzy parameterization and TSK-style rule reasoning are employed.
• A rule-constrained fuzzy graph attention encoder modulates neighborhood aggregation to reduce noise, and a pluggable membership injection strategy enhances decoder-agnosticism.

Main Results:

• The proposed framework effectively suppresses uncertainty-induced noise accumulation through reliability-modulated neighborhood aggregation.
• Experiments demonstrate the effectiveness, robustness, and adaptability of FGF-GAT across entity and relation prediction tasks.
• The framework's decoder-agnostic nature is confirmed by successful integration with various scoring functions.

Conclusions:

• FGF-GAT provides a general and effective solution for fine-grained fuzzy KGE, enhancing reasoning with uncertain information.
• The framework's design improves interpretability and reduces noise, paving the way for more reliable AI decision-making.